Protein function is predictable from a taxonomy of orthogonal evidence, combined by a calibrated fusion, measured on a leakage-free temporal frame. The nine-cell benchmark is not nine numbers, it is a map of which evidence wins in which regime, and its frontier is a measured limit of ranking rather than of evidence, which this work characterises rather than hides.

What this is. A protein is a chain of amino acids that does a job in the cell. Its function is that job, written in a shared vocabulary called the Gene Ontology, or GO: terms like binds DNA or located in the mitochondrion. Most proteins ever sequenced carry no such labels. PROTEA reads a protein's sequence and proposes its GO terms, with the evidence for each one in plain view.

How it works, in one breath. PROTEA does not guess a function from the sequence directly. It retrieves. It turns each protein into a compact code, finds the most similar proteins that are already labelled, and lets those neighbours vote for GO terms; this is k-nearest-neighbours, kNN. A second stage, a reranker, weighs that vote against other clues, how strong each match is and how a candidate term relates to what we already know, and calibrates a final score.

What it achieves. On a fair test, PROTEA ranks first in seven of the nine evaluation cells on the field's headline score (a weighted, information-aware measure called f_micro_w). The figure itself is withheld while the campaign recomputes it, and the ranking is read from the sealed board, which has not moved. The nine cells are three knowledge regimes, from proteins we know nothing about to proteins we already know something about, crossed with the three branches of GO.

Why the test is fair. The hard part of this field is not the model, it is not cheating. PROTEA is scored on a temporal holdout: we freeze what was known on one date and ask only about function that was discovered afterwards, so the answer cannot leak into the question. Every number here is reproducible from that frozen frame, and the sealed board never moves; regenerated numbers are candidates until reviewed against it.

Where it stops, honestly. Two of the nine cells are not won: predicting the Biological Process branch for the least-studied proteins. We call it the biological-process wall, we show it rather than hide it, and we measured where it lives. It is not that the evidence is missing: almost all of those terms are already visible to the method. It is that we cannot order the shortlist we build. A perfect ranking of the very candidates we already retrieve would score close to three times what we deliver.